feat: 增强任务步骤注入与a11y状态表达，提升树形交互稳定性

- 打通 metadata.steps 传递链路，将任务步骤注入 agent 预测上下文

- 优化 a11y tree 线性化输出：使用中心坐标并新增 states 列（expanded/collapsed/selected 等）

- 放宽可保留节点条件，保留无文本输入类控件（edit/textfield/searchbox 等）

- 强化输出约束：单轮仅允许动作代码或 WAIT/DONE/FAIL，禁止动作与 DONE 同轮返回

- 补充 avogadro 示例步骤：展开 aromatics 并选择 benzene.cjson

lib_run_single.py

@@ -9,7 +9,7 @@ from lib_results_logger import log_task_completion
 logger = logging.getLogger("desktopenv.experiment")
 
 
-def run_single_example(agent, env, example, max_steps, instruction, args, example_result_dir, scores):
+def run_single_example(agent, env, example, max_steps, instruction, args, example_result_dir, scores, metadata_steps=""):
     runtime_logger = setup_logger(example, example_result_dir)
 
     # Reset environment first to get fresh VM IP
@@ -36,7 +36,8 @@ def run_single_example(agent, env, example, max_steps, instruction, args, exampl
         logger.info(f"Step {step_idx + 1} prediction...")
         response, actions = agent.predict(
             instruction,
-            obs
+            obs,
+            metadata_steps=metadata_steps,
         )
         logger.info(f"Response: {response}")
         logger.info(f"Actions: {actions}")